What's New
Smart Dispatch
get_dispatch_strategy()analyzes task type to choose 2 vs 3 providers- Research: 2 providers (Gemini + Claude-sonnet) — OctoBench showed marginal ROI from 3rd
- Review/Security: 3 providers — each misses different bugs (proven by benchmark)
- Architecture/Debate: 2 providers (Codex + Gemini for max training bias diversity)
- Override:
OCTOPUS_DISPATCH_STRATEGY=full|smart|minimal
Blind Spot Library
- 25 curated blind spots across 5 domains (security, architecture, data, API, operations)
- Perspectives LLMs systematically miss, sourced from OctoBench benchmark data
- Auto-injected into edge-case and synthesis perspectives during research
- Two-tier keyword matching: manifest triggers → per-spot triggers
Researcher Persona Improvements
- Mandatory balance requirement: argue BOTH sides of recommendations
- Compliance/regulatory awareness for enterprise/B2B contexts
- Gemini-first probe ordering (strongest research breadth per benchmark)
Bug Fix: Skill Tool References
- Fixed 16 commands using wrong skill names (
octo:X→ actual skill names likeflow-discover) - This caused "Unknown skill" errors on every
/octo:command invocation